Cybersecurity News: OpenAI takes on Mythos, Klue hits security shops, Five Eyes has eyes on AI

By
Sarah Lane
-

In today’s cybersecurity news…

OpenAI takes on Anthropic’s Mythos

OpenAI is rolling out GPT-5.5 Cyber, its new cybersecurity-focused model, and kicking off a project called Patch the Planet to help open source developers find and fix security bugs. OpenAI says the updated model now outperforms Anthropic’s Mythos 5 on a leading cybersecurity test, while the new program has already uncovered hundreds of vulnerabilities and produced dozens of patches. The move comes as AI companies face growing scrutiny over how powerful cyber-focused models are becoming, and how they’ll be used.  (Wired)

Klue hack hits security shops

More companies are coming forward as victims of the Klue breach, including security firms like Huntress, Recorded Future, Tanium, HackerOne, and Snyk. The attackers accessed Salesforce-related sales and business data through a compromised Klue integration, but companies say passwords, threat intelligence, and core product systems weren’t affected. The campaign is being linked to a newer extortion group called Icarus, which is now threatening to leak data from what could be hundreds of organizations. (The Register)

Five Eyes has eyes on AI models

The Five Eyes intelligence alliance is warning that AI models capable of dramatically changing cyberattacks and cyber defense could arrive within months, instead of years. The unusual public warning comes just weeks after the US administration blocked foreign nationals from accessing Anthropic’s Fable model over cybersecurity concerns. Experts say the bigger issue is that even if Fable is restricted, plenty of other companies and countries are racing to build similarly powerful systems. (The Guardian)

Microsoft fixes AutoGen Studio flaw

Microsoft fixed a vulnerability called AutoJack in its open source AutoGen Studio platform that could have let attackers trick an AI agent into running commands on a developer’s machine simply by getting it to visit a malicious webpage. The bug combined several authentication and WebSocket flaws that could be used to launch PowerShell, Bash, or other commands with the user’s privileges. The good news is it never made it into an official release. But developers who built AutoGen Studio directly from GitHub during a brief development window may have been exposed.  (BleepingComputer)

Huge thanks to our sponsor, Guardsquare

Your backend is only as secure as your frontend. Research shows that client-side compromise is now a primary driver of API risk. With sixty-three percent of leaders detecting mobile app tampering or cloning last year, don’t leave your mobile app security to chance. Get multilayered protection for your entire mobile app ecosystem from the outside in. Learn more at Guardsquare.com.

DifyTap flaws could expose AI chats

Researchers at Zafran Security found four vulnerabilities in the popular open source AI platform Dify that could have let attackers secretly access private AI chats and uploaded files belonging to other customers. Some of the flaws didn’t require authentication and could be abused across different tenants, siphoning off conversations and AI responses from victim applications. Most of the bugs have already been patched, but it’s another reminder that small authorization mistakes can create big security problems. (The Hacker News)

Meta’s employee-tracking program releases data

WIRED cited an internal security notice and three current Meta employees speaking to the publication who say Meta has paused its employee-monitoring program after discovering data collected from workers’ laptops was left accessible internally due to a misconfiguration. The program was gathering things like keystrokes, mouse clicks, screen activity, AI prompts, private conversations, and other employee data to help train AI models, and more than 1,600 employees had already protested it over privacy and security concerns. (Wired)

Cloudflare teams up with browsers to sniff out bots

Cloudflare is teaming up with Google, Microsoft, and Mozilla on a new system called Private Access Control Tokens, or PACTs, that would let websites tell the difference between legitimate users and unwanted bot traffic without constantly throwing CAPTCHAs at people. Think of it as a privacy-preserving trust signal that browsers, and eventually AI agents, could present to websites to prove they’re behaving legitimately. The goal is to cut down on fraud and abusive traffic. (The Register)

FFmpeg fixes PixelSmash flaw in video decoder

JFrog researchers discovered a high-severity FFmpeg vulnerability called PixelSmash that could let attackers crash applications or even execute code through a malicious video file. The flaw affects FFmpeg’s widely used MagicYUV decoder and could impact apps and services that rely on FFmpeg, including Jellyfin, Kodi, OBS Studio, PhotoPrism, and potentially messaging platforms that generate video previews. FFmpeg appeared to fix the bug in version 8.1.2. (BleepingComputer)